---
title: The 3-Options Problem — Why AI Curation Beats Search
description: Showing 500 flight results is not helpful — it is hostile. AI curation narrows to 3 ranked options and explains why each one matters.
canonical: https://nowah.xyz/blog/three-options-problem-ai-curation
lastModified: "2026-08-07T08:08:37.847Z"
---

# The 3-Options Problem — Why AI Curation Beats Search

Showing 500 flight results is not helpful — it is hostile. AI curation narrows to 3 ranked options and explains why each one matters.

Here is a number that should embarrass every travel platform on the market: 67% of travelers find the booking process stressful. Not the travel. The booking. They are stressed by the act of purchasing something they are excited about.

The root cause is information overload. A typical flight search returns 200 to 500 results. The platform throws all of them at you with a sort button and a prayer. Good luck figuring out which one is actually best for you.

Barry Schwartz described this phenomenon in The Paradox of Choice: more options do not make people happier. Past a threshold, more options make people anxious, indecisive, and less satisfied with whatever they eventually choose. Travel booking is the most extreme consumer example of this paradox in action.

We decided to take a different approach at Nowah. Instead of showing every available option and making the user sort through them, our AI agent curates exactly three options and explains why each one was selected. I want to walk through why three, how the ranking pipeline works, and what the data says about this approach versus traditional search.

## The 500-result problem

![Illustration for this section](https://pics.nowah.xyz/website-media/ai-research-004-img-1.webp)

When you search for flights on a traditional platform, the search engine queries global inventory and returns everything that matches your basic parameters. Departing from SFO, arriving at NRT, sometime in April, economy class. That query might return 400 results.

The platform gives you tools to narrow this down: sort by price, filter by number of stops, set departure time ranges, exclude specific airlines. But these tools put the entire cognitive burden on you. You have to figure out which filters matter, apply them in the right combination, and then manually compare the remaining options.

Price comparison alone takes over 3 hours for a single round-trip flight when done across platforms. And price is one of at least 20 factors that determine whether a flight is actually good for you.

The result is a paradox. The platform optimized for showing you maximum choice. But maximum choice does not lead to maximum satisfaction. It leads to decision paralysis, satisficing (picking something "good enough" out of exhaustion), and post-booking regret. About 40% of travelers report wishing they had found a better option after booking.

## Inside the ranking pipeline

Our agent takes a fundamentally different approach. When you ask for flights to Tokyo, here is what happens behind the scenes.

**Step 1: Broad search.** The agent queries live flight inventory with your parameters. This returns the full result set, often hundreds of options. The user never sees this raw list.

**Step 2: Hard filtering.** Options that violate firm constraints get eliminated immediately. If you said "no red-eyes," every overnight flight is removed. If your maximum budget is $1,500, anything above is gone. If the agent knows from your history that you avoid a specific airline, those flights are filtered out. Hard filters typically eliminate 50-60% of results.

**Step 3: Soft scoring.** The remaining options get scored on a multi-factor rubric. This is where it gets interesting. The agent does not just sort by price. It evaluates:

- Total price (including baggage fees and fare class restrictions)
- Number of stops and total travel time
- [Connection quality](/blog/airport-intelligence-connection-quality) (is a 90-minute layover in a sprawling international terminal actually doable?)
- Departure and arrival time alignment with your preferences
- Airline quality and your history with that carrier
- Seat availability in your preferred cabin class
- Fare flexibility (can you change it if plans shift?)

Each factor gets a weight. Those weights are partially universal (everyone prefers shorter travel time) and partially personal (you care about connection quality more than most because you have tight connections in your history). The agent calibrates weights based on what it knows about you.

**Step 4: Diversity check.** After scoring, we do not just take the top three by score. We ensure the three options represent genuine variety. Typically, we present one "best overall" option that scores highest on the composite metric, one "best value" option that offers the strongest price-to-quality ratio, and one "[best experience](/blog/best-experience-booking-ai-curates)" option that maximizes comfort and convenience.

This diversity check is important because it gives the user meaningful choice without overwhelming them. Three options that are nearly identical are not useful. Three options that each represent a different tradeoff let the user decide which tradeoff they care about.

**Step 5: Narrative explanation.** The agent does not just present three cards. It explains why each was selected. "Option A is the direct morning flight you typically prefer, at $520. Option B saves $140 but adds a 2-hour layover in Seoul, which I flagged because Incheon is a comfortable airport and 2 hours is plenty. Option C is $80 more than A but includes a premium economy upgrade with better seat pitch."

This explanation is not decoration. Trust in AI recommendations jumps from roughly 30% to roughly 65% when the AI explains its reasoning. Showing the work is the single biggest lever for user confidence.

## Why three and not five or ten

![Supporting diagram](https://pics.nowah.xyz/website-media/ai-research-004-img-2.webp)

We tested different numbers. One option feels dictatorial. Users want choice, even when too much choice is harmful. Two options feel like a binary trade-off and users worry they are missing something. Five or more options start to recreate the overwhelm problem.

Three is the sweet spot. It is small enough that the user can fully evaluate each option without fatigue. It is large enough to represent genuine variety. And it maps naturally to the "good, better, best" mental model that people use intuitively.

We also give users the option to see more. "Show me more options" expands to five or lets the user specify new criteria. But the default is three, and the majority of bookings come from the initial set.

## A/B evidence

The data on curated versus uncurated results is clear.

Users shown 3 AI-curated options report higher satisfaction with their booking than users shown 50+ options and asked to filter. This seems counterintuitive until you think about what satisfaction actually means. It is not "did I see the absolute cheapest option?" It is "do I feel confident I made a good choice?"

Decision time drops significantly with curated results. When you are comparing 3 options with clear explanations of trade-offs, you can make a decision in under a minute. When you are scrolling through 200 results trying to figure out if the $20 cheaper option with a worse layover is worth it, you can burn 30 minutes.

Post-booking regret, measured by how often users search for the same route after booking, is lower with curated results. This makes sense: the agent already evaluated the alternatives, so the user is less likely to wonder "what if I missed something?"

Conversion rate is higher because the path from "here are your options" to "book this one" is dramatically shorter. The traditional funnel has dozens of drop-off points. The curated funnel has one decision point.

## When three is not enough

There are legitimate cases where three options do not cover the space.

**Complex multi-city trips** might need different top-three sets for each leg. The agent handles this by presenting options per segment while highlighting how they fit together.

**Group travel with \[conflicting preferences\]\(/blog/how\-ai\-handles\-conflicting\-preferences\)** might require showing how different options serve different people. "Option A is best for your budget, Option B is best for your partner's schedule preference."

**Highly price-sensitive searches** where the user explicitly wants to see the full range. Some users are experienced bargain hunters who enjoy the search process. We let them opt into a broader view, but the default remains curated.

**Unusual routes or dates** where there genuinely are not many options. If only five flights operate on your route and date, showing three curated options from five total does not add much value. In these cases, the agent transparently says "there are only five options on this route, here they are" and presents all of them.

The key is that curation is the default, and expansion requires an explicit request. This is the opposite of how traditional platforms work, where the default is overwhelm and narrowing requires effort.

## The narrative layer: why explanation matters

Presenting three options is half the job. Explaining why each was selected is the other half.

When the agent presents three flights without explanation, the user is in the same position as before: evaluating options based on visible data (price, times, stops) and making their own judgment. The curation happened behind the scenes, but the user cannot evaluate whether it was good curation or arbitrary selection.

When the agent explains its reasoning, everything changes. "Option A is the direct morning flight you typically prefer. Option B saves $140 with a comfortable 2-hour layover at a well-rated airport. Option C is $80 more but includes premium economy with extra legroom, which I flagged because you mentioned knee discomfort on your last long-haul flight."

That narrative does several things simultaneously. It proves the agent knows the user (knee discomfort, morning preference). It demonstrates that each option was selected for a specific reason, not randomly. It articulates trade-offs that the user might not have considered ([layover quality](/blog/data-behind-layover-quality), legroom value). And it gives the user a framework for deciding.

Trust in AI recommendations jumps from roughly 30% to roughly 65% when the AI explains its reasoning. That is the difference between a product people use cautiously and a product people rely on. The narrative layer is the trust layer.

We experimented with different levels of explanation detail. Too brief ("here are three good options") does not build trust. Too verbose (paragraph-long explanations for each option) is not read. The sweet spot is 1-2 sentences per option that capture the core trade-off and the personal relevance.

The narrative also needs to include a recommendation. Users want the agent's opinion. "If I were you, I would go with Option A because the direct flight saves 4 hours of travel time and the price is reasonable for this route" is more helpful than neutrally presenting three options. The recommendation can be wrong. But having an opinion is what separates a capable assistant from a search engine.

## How personalization improves curation over time

The first time you use Nowah, the agent curates based on general population preferences. Most people prefer direct flights. Most people prefer reasonable departure times. Most people weight price heavily.

After your first trip, the curation adjusts. You picked the connection option over the direct flight, signaling that you are more price-sensitive than average. Or you picked the more expensive option with better timing, signaling that convenience outweighs price for you.

After five trips, the curation is meaningfully personal. The agent knows your price sensitivity, your time-of-day preferences, your airline comfort levels, your layover tolerance, and your cabin class flexibility. The three options it presents are not "three good flights for anyone." They are "three good flights for you specifically."

After ten trips, the top recommendation is often the one you would have chosen. The confirmation accept rate rises because the agent's model of your preferences converges with your actual decision-making patterns.

This progression is measurable. We track preference match scores over time and see a clear logarithmic improvement curve. The biggest gains happen in the first 3-5 trips. By trip 10, the agent's recommendations are well-calibrated. This is the compound effect of memory plus curation: each trip teaches the agent something that makes the next curation better.

Price comparison across platforms takes over 3 hours for a single round-trip flight. After a few trips with Nowah, the agent already knows what "good" looks like for you. The 3-hour comparison is reduced to a 3-second curation. That is the value proposition.

## When 3 is not enough

The 3-options model is a default, not a rule. Some situations demand more.

**Flexible travelers** who genuinely enjoy browsing want to see a wider range. "Show me everything under $600" is a legitimate request. The agent respects it by presenting more options, organized by category rather than dumped in a raw list. "Here are 8 options under $600: 3 direct flights, 3 with short connections, and 2 budget options with longer layovers." The curation still adds structure even when the count increases.

**High\-\[stakes decisions\]\(/blog/ai\-safety\-high\-stakes\-decisions\)** sometimes need more comparison points. A honeymoon trip where the user is deliberating between business class and premium economy might benefit from 5-6 options that span the spectrum. The agent reads the situation and adjusts the count.

**Multi-leg trips** are inherently more complex. For a 3-city itinerary, showing 3 options for each leg means 9 total flight selections. The agent might present the top 2 per leg to keep the decision manageable, or bundle them into complete itinerary packages where each package represents a coherent set of choices.

The important principle is not "always show 3." It is "curate to the minimum number that gives the user genuine choice without creating overload." Sometimes that number is 3. Sometimes it is 5. Rarely is it 50.

We track what we call the "expansion rate": how often users ask to see more options beyond the initial curation. An expansion rate above 30% suggests we are under-curating, showing too few or the wrong options. Below 10% suggests the initial curation is well-calibrated. We currently sit around 15-20%, which feels right. Some users naturally want to explore, and the agent should support that without making exploration the default.

## The ranking pipeline in detail

Behind the 3 options the user sees is a multi-stage pipeline that processes hundreds of raw results.

**Stage 1: Hard filters.** These are binary constraints from the user's request. Dates, airports, number of stops, cabin class. A user who said "direct flights only" eliminates every option with a connection. Hard filters typically reduce 500 results to 100-200.

**Stage 2: Soft scoring.** Each remaining option gets a composite score based on multiple dimensions. Price (normalized against the route average). Duration (shorter is better, with diminishing returns). Departure time (aligned with the user's preference or general population preference if unknown). Airline quality (based on on-[time performance](/blog/airline-on-time-performance-who-delivers) and service ratings). Connection quality (for non-direct flights: layover duration, airport quality, terminal distance).

Each dimension has a weight. Weights are personalized based on the user's history. A user who consistently picks the cheapest option gets a higher price weight. A user who consistently picks direct flights gets a higher stop-count weight. For new users, we use population-level weights that reflect average preferences.

**Stage 3: Diversity check.** The top 3 by composite score might all be similar: three direct flights on different airlines at similar prices. This is technically optimal but does not give the user genuine choice. The diversity check ensures the final 3 represent different trade-off profiles. Typically: best overall (highest composite score), best value (strong on price dimension), and best experience (strong on comfort and timing dimensions).

**Stage 4: Narrative generation.** Each selected option gets a 1-2 sentence explanation connecting it to the user's specific situation. This is where the ranking becomes visible to the user.

The entire pipeline runs in seconds. The user sees 3 options with explanations. Behind those options is a sophisticated multi-stage process that considered hundreds of alternatives and selected the most useful subset.

## The deeper argument

I think the 3-options model represents something bigger than a UX improvement. It reflects a shift in what travel technology is for.

Traditional platforms are inventory displays. Their job is to show you what is available. The intellectual work of evaluating options is entirely yours.

AI curation means the technology does the evaluation. It considers factors you do not have time to research, applies preferences you do not have to re-state, and presents a small set of well-chosen options with clear reasoning.

The difference matters at the unit economics level too. A platform that displays 500 results needs the user to spend 45 minutes evaluating them. An AI agent that curates 3 options needs the user to spend 3 minutes deciding. The faster path to booking means higher conversion, lower abandonment, and a fundamentally different relationship with the user's time.

Sixty-seven percent of travelers find booking stressful. The stress comes from information overload, not from lack of information. Curation is the antidote to overload.

This is how a good human travel agent works. You do not call a travel agent and ask them to read you a list of every available flight. You tell them what you want and they come back with a recommendation. Maybe two or three options with different tradeoffs. They explain the reasoning. You decide.

## What curation means for the travel industry

The shift from search to curation has implications beyond individual products. It changes the competitive dynamics of the entire travel industry.

Platforms that display hundreds of results can monetize through advertising: sponsored listings, banner ads, affiliate placement. When curation reduces the display to 3 options, the advertising surface area vanishes. You cannot run sponsored listings when there are only 3 results and the user trusts the agent to select them honestly.

This means AI-curated travel must monetize differently. Transaction fees. Subscription models. Value-added services. The revenue model follows from the product model, and the product model is curation, not advertising.

For consumers, this is better. The incentives align. The agent's goal is to find you the best option, not to display the option that paid for placement. Trust in the recommendations increases because the business model does not depend on biasing them.

For the industry, it is a disruption. The platforms built on advertising-supported search have to find new revenue models or lose share to AI-native competitors. This transition is happening now, and it favors companies that built for curation from day one rather than retrofitting it onto an advertising model.

The best travel app is the one that takes the work off your plate. Three well-chosen options with honest explanations. That is the product. Everything else is noise.

---

Nowah is an AI travel agent that searches and books real flights and hotels through conversation — no filters, no thirty open tabs. [Plan your next trip](https://app.nowah.xyz).